Arizona State University's Narrative and Emerging Media Program has detailed Open4DGS, a pipeline that turns multi-camera footage from consumer cameras into 4D Gaussian splatting sequences.
Open4DGS is built for rigs without hardware genlock, assembled from rolling shutter cameras such as GoPros, Sony action cameras or mixed consumer setups. The project page puts a genlocked array of broadcast global shutter cameras at tens of thousands of dollars. Backblaze's open source 4D splatting template, published in October, starts from video that is already synchronized.
The pipeline runs in seven stages. Frames are extracted and aligned by timecode, cameras are calibrated with COLMAP, and foregrounds are masked with Background Matting V2 or SAM 2. A solver script, calculate-sync.py, then estimates each camera's sub-frame timing offset and rolling shutter readout by tracking epipolar flow across calibrated camera pairs, with an example readout of 15.0 ms for a Sony RX0 II at 60 fps. When its closed loop consistency checks fail, the solver falls back to timecode alone.
Open4DGS matches points with RoMa v2 across cameras and MEMFOF across time. Points moving faster than 4.0 spacings per frame trigger one frame sub-keyframes, and training gives those fast frames additional visits.
The training core builds on TRiGS, a method Yeom et al. posted to arXiv in April, and its reference code TriGSVanilla. In Open4DGS, each primitive's translation and rotation are optimized jointly as a single twist, with quadratic Bézier control points modeling curved paths. Open4DGS fixes the canonical primitive count and recycles low opacity slots into active regions, so GPU memory stays constant, and it splits long captures into temporal chunks. Output is a .trigsbin timeline plus per frame PLY files for Blender, Unreal Engine and WebGL viewers, and a viewer shows the reconstruction while it trains.
The Open4DGS project page credits the work to the NEM Program, part of ASU’s Sidney Poitier New American Film School, and lists Arvin Tehrani of Contrast Media Labs, Rory Mitchell, Ryat Yezbick and Nonny de la Peña as its authors. The open source release was driven by de la Peña, the program’s founding director, who bought a capture rig for ASU’s Narrative and Emerging Media program and hired Contrast Media Labs to write the open source code to help advance the industry and community.
The page also credits the open source work Open4DGS builds on, naming TriGSVanilla and TRiGS, RoMa v2, MEMFOF from the MSU Video Group, SAM 2 from Meta, Background Matting V2, gsplat and COLMAP. The software is licensed under MIT and AGPL-3.0.
The code is split across two repositories. Open4DGS-Core prepares synchronized multi-camera video, initializes dynamic Gaussian scenes and exports TRiGS reconstructions, and training lives in the companion Open4DGS-Trigs repository under AGPL-3.0. The main Linux setup was exercised on Ubuntu 24.04 with Python 3.11, CUDA 12.8 and an NVIDIA RTX 3090, a Windows 11 configuration on an RTX PRO 6000 Blackwell is reported, and WSL 2 has not yet been validated.
Two example datasets are planned for a separate Open4DGS-Datasets repository, and the developers told Radiance Fields they will carry a research and education license. The roadmap, which carries no committed dates, lists prep and training time improvements, a Unity plugin with a sample project, web SDKs based on Spark and three.js, an 8th Wall sample project, and compression and streaming.
The pipeline is detailed on the Open4DGS project page.




